ITK学习笔记(八) ITK高斯混合模型 GMM EM
1、高斯混合模型
sklearn.mixture
是一个能够学习高斯混合模型、抽样高斯模型和从数据中估计模型的包。同样,也提供了帮助决定正确组件数量的方法。
一个高斯混合模型是一个概率模型,它假设所有的数据点是从有限未知参数的高斯分布的混合生成的。可以将混合模型当作泛化的k均值聚类,以融合关于数据协方差和潜在高斯中心的信息。
高斯混合
GaussianMixture
对象实现了expection-maximization
算法来拟合高斯混合模型。它也能够得到多元模型的置信椭圆,计算贝叶斯信息准则来确定数据中聚集类别的数量。GaussianMixture.fit
方法从训练数据中学习一个高斯混合模型,GaussianMixture.predict
能够分配给每个样本最大可能属于的高斯分布。
GaussianMixture
提供了不同的选项来限制不同类别估计的方差,包括,spherical、diagonal、tied或full方差。
2、变分贝叶斯高斯混合
BayesianGaussianMixture
对象实现了一系列考虑不同推断算法的高斯混合模型。
估计算法:变分推断
变分推断(Variational Inference
)是最大期望的扩展,它最大化模型证据的下界,而不是数据似然。其背后的原理与最大期望方法相同。但是变分推断方法通过集成先验分布的信息添加正则项。这可以避免在最大期望中经常发生的奇异性,但会引入偏差到模型中。
BayesianGaussianMixture
类提供了两类权重的先验:使用Dirichlet
分布的有限混合模型和使用Dirichlet
过程的无限混合模型。
参考:高斯混合模型
3、ITK中的GMM、EM
使用 ITK中的GMM、EM进行分布式采样
#include "itkVector.h"
#include "itkListSample.h"
#include "itkGaussianMixtureModelComponent.h"
#include "itkExpectationMaximizationMixtureModelEstimator.h"
#include "itkNormalVariateGenerator.h"
int
main(int, char *[])
{
unsigned int numberOfClasses = 2;
using MeasurementVectorType = itk::Vector<double, 1>;
using SampleType = itk::Statistics::ListSample<MeasurementVectorType>;
SampleType::Pointer sample = SampleType::New();
using NormalGeneratorType = itk::Statistics::NormalVariateGenerator;
NormalGeneratorType::Pointer normalGenerator = NormalGeneratorType::New();
normalGenerator->Initialize(101);
MeasurementVectorType mv;
double mean = 100;
double standardDeviation = 30;
for (unsigned int i = 0; i < 10; ++i)
{
mv[0] = (normalGenerator->GetVariate() * standardDeviation) + mean;
std::cout << "m[" << i << "] = " << mv[0] << std::endl;
sample->PushBack(mv);
}
normalGenerator->Initialize(3024);
mean = 200;
standardDeviation = 30;
for (unsigned int i = 0; i < 10; ++i)
{
mv[0] = (normalGenerator->GetVariate() * standardDeviation) + mean;
std::cout << "m[" << i << "] = " << mv[0] << std::endl;
sample->PushBack(mv);
}
using ParametersType = itk::Array<double>;
ParametersType params1(2);
std::vector<ParametersType> initialParameters(numberOfClasses);
params1[0] = 110.0;
params1[1] = 50.0;
initialParameters[0] = params1;
ParametersType params2(2);
params2[0] = 210.0;
params2[1] = 50.0;
initialParameters[1] = params2;
using ComponentType = itk::Statistics::GaussianMixtureModelComponent<SampleType>;
std::vector<ComponentType::Pointer> components;
for (unsigned int i = 0; i < numberOfClasses; i++)
{
components.push_back(ComponentType::New());
components[i]->SetSample(sample);
components[i]->SetParameters(initialParameters[i]);
}
using EstimatorType = itk::Statistics::ExpectationMaximizationMixtureModelEstimator<SampleType>;
EstimatorType::Pointer estimator = EstimatorType::New();
estimator->SetSample(sample);
estimator->SetMaximumIteration(500);
itk::Array<double> initialProportions(numberOfClasses);
initialProportions[0] = 0.5;
initialProportions[1] = 0.5;
estimator->SetInitialProportions(initialProportions);
for (unsigned int i = 0; i < numberOfClasses; i++)
{
estimator->AddComponent((ComponentType::Superclass *)components[i].GetPointer());
}
estimator->Update();
for (unsigned int i = 0; i < numberOfClasses; i++)
{
std::cout << "Cluster[" << i << "]" << std::endl;
std::cout << " Parameters:" << std::endl;
std::cout << " " << components[i]->GetFullParameters() << std::endl;
std::cout << " Proportion: ";
std::cout << " " << estimator->GetProportions()[i] << std::endl;
}
return EXIT_SUCCESS;
}
运行输出结果:
前面20行分别是生成的2类,各10个样本,1类样本均值100,方差30;2类样本均值200,方差30.
21~24行是分类结果的1类参数;
25~28行是分类结果的2类参数;
占比之和为1;
参数应该是均值和方差,均值比较接近生成时的值 100和200,方差与原来的30差别巨大,还不知道怎么理解。
m[0] = 156.311
m[1] = 205.464
m[2] = 80.8426
m[3] = 136.952
m[4] = 86.6091
m[5] = 80.3185
m[6] = 107.911
m[7] = 63.1748
m[8] = 107.082
m[9] = 112.343
m[0] = 189.946
m[1] = 174.951
m[2] = 243.387
m[3] = 169.488
m[4] = 261.163
m[5] = 215.278
m[6] = 212.506
m[7] = 150.613
m[8] = 186.132
m[9] = 213.155
Cluster[0]
Parameters:
[91.04822175454494, 385.98395103056583]
Proportion: 0.325826
Cluster[1]
Parameters:
[189.88473393439773, 1626.5175226586712]
Proportion: 0.674174
后来在代码中添加了打印信息:
estimator->Print(std::cout);
打印输出:
ExpectationMaximizationMixtureModelEstimator (0000015D6BE586B0)
RTTI typeinfo: class itk::Statistics::ExpectationMaximizationMixtureModelEstimator<class itk::Statistics::ListSample<class itk::Vector<double,1> > >
Reference Count: 1
Modified Time: 74
Debug: Off
Object Name:
Observers:
none
Maximum Iteration: 100
Sample: 0000015D6BE2B790
Number Of Components: 2
Component Membership Function[0]: 0000015D6BE53FC0
Component Membership Function[1]: 0000015D6BE58540
Termination Code: itk::Statistics::ExpectationMaximizationMixtureModelEstimatorEnums::TERMINATION_CODE::NOT_CONVERGED
Initial Proportions: [0.5, 0.5]
Proportions: [0.3258255562923341, 0.6741744437076659]
Calculated Expectation: -15.0301
从打印结果来看,EM算法更新了74次参数,设置的最大迭代次数是100,因此迭代次数是够用的,最终收敛期望是 -15.03, 退出代码说明 TERMINATION_CODE::NOT_CONVERGED 未收敛。
这是官方示例程序,所以不知道说啥。
正常理解应该是收敛到0附近。